diff --git a/a.out b/a.out
index 9c30d5e..b6e3423 100755
Binary files a/a.out and b/a.out differ
diff --git a/build/a.asm b/build/a.asm
index bf532f2..2cc2435 100644
--- a/build/a.asm
+++ b/build/a.asm
@@ -18,17 +18,30 @@ _start:
     mov rdi, rax
     mov rax, 60
     syscall
+word_c_40:
+	mov rax, [r12]
+	movzx rax, byte [rax]
+	mov [r12], rax
+	ret
+    ret
+word_c_21:
+	mov rax, [r12]
+	add r12, 8
+	mov rbx, [r12]
+	mov [rbx], al
+	ret
+    ret
+word_r_40:
+	mov rax, [r13]
+	sub r12, 8
+	mov [r12], rax
+	ret
+    ret
 word_strlen:
-	mov rsi, [r12]      ; addr
-	mov rcx, 0
-strlen_loop:
-	mov al, [rsi + rcx]
-	cmp al, 0
-	je strlen_done
-	inc rcx
-	jmp strlen_loop
-strlen_done:
-	mov [r12], rcx      ; len
+	mov rax, [r12]      ; addr
+	mov rcx, [r12 + 8]  ; len
+	add r12, 16         ; pop len and addr
+	mov [r12], rcx      ; push len
 	ret
     ret
 word_puts:
@@ -113,6 +126,15 @@ word_dup:
 	sub r12, 8
 	mov [r12], rax
     ret
+word_write_buf:
+	mov rdx, [r12]        ; len
+	mov rsi, [r12 + 8]    ; addr
+	add r12, 16           ; pop len + addr
+	mov rax, 1            ; syscall: write
+	mov rdi, 1            ; fd = stdout
+	syscall
+	ret
+    ret
 word_drop:
 	add r12, 8
     ret
@@ -284,20 +306,22 @@ word__21:
 	add r12, 8
     ret
 word_mmap:
-	mov r9, [r12]
-	add r12, 8
-	mov r8, [r12]
-	add r12, 8
-	mov r10, [r12]
-	add r12, 8
-	mov rdx, [r12]
-	add r12, 8
-	mov rsi, [r12]
-	add r12, 8
-	mov rdi, [r12]
-	mov rax, 9
-	syscall
-	mov [r12], rax
+    ; Save rsp and align to 16 bytes for syscall ABI
+    mov rax, rsp
+    and rsp, -16
+    mov rdi, [r12+40]   ; addr
+    mov rsi, [r12+32]   ; length
+    mov rdx, [r12+24]   ; prot
+    mov r10, [r12+16]   ; flags
+    mov r8,  [r12+8]    ; fd
+    mov r9,  [r12]      ; offset
+    add r12, 48         ; pop 6 args
+    mov rax, 9          ; syscall: mmap
+    syscall
+    mov rsp, rax        ; restore rsp
+    sub r12, 8
+    mov [r12], rax      ; push result
+    ret
     ret
 word_munmap:
 	mov rsi, [r12]
diff --git a/build/a.o b/build/a.o
index 6e22e17..f77941f 100644
Binary files a/build/a.o and b/build/a.o differ
diff --git a/build/hello.asm b/build/hello.asm
index 12c9dff..2cc2435 100644
--- a/build/hello.asm
+++ b/build/hello.asm
@@ -18,6 +18,32 @@ _start:
     mov rdi, rax
     mov rax, 60
     syscall
+word_c_40:
+	mov rax, [r12]
+	movzx rax, byte [rax]
+	mov [r12], rax
+	ret
+    ret
+word_c_21:
+	mov rax, [r12]
+	add r12, 8
+	mov rbx, [r12]
+	mov [rbx], al
+	ret
+    ret
+word_r_40:
+	mov rax, [r13]
+	sub r12, 8
+	mov [r12], rax
+	ret
+    ret
+word_strlen:
+	mov rax, [r12]      ; addr
+	mov rcx, [r12 + 8]  ; len
+	add r12, 16         ; pop len and addr
+	mov [r12], rcx      ; push len
+	ret
+    ret
 word_puts:
 	; detects string if top is len>=0 and next is a pointer in [data_start, data_end]
 	mov rax, [r12]      ; len or int value
@@ -100,6 +126,15 @@ word_dup:
 	sub r12, 8
 	mov [r12], rax
     ret
+word_write_buf:
+	mov rdx, [r12]        ; len
+	mov rsi, [r12 + 8]    ; addr
+	add r12, 16           ; pop len + addr
+	mov rax, 1            ; syscall: write
+	mov rdi, 1            ; fd = stdout
+	syscall
+	ret
+    ret
 word_drop:
 	add r12, 8
     ret
@@ -271,20 +306,22 @@ word__21:
 	add r12, 8
     ret
 word_mmap:
-	mov r9, [r12]
-	add r12, 8
-	mov r8, [r12]
-	add r12, 8
-	mov r10, [r12]
-	add r12, 8
-	mov rdx, [r12]
-	add r12, 8
-	mov rsi, [r12]
-	add r12, 8
-	mov rdi, [r12]
-	mov rax, 9
-	syscall
-	mov [r12], rax
+    ; Save rsp and align to 16 bytes for syscall ABI
+    mov rax, rsp
+    and rsp, -16
+    mov rdi, [r12+40]   ; addr
+    mov rsi, [r12+32]   ; length
+    mov rdx, [r12+24]   ; prot
+    mov r10, [r12+16]   ; flags
+    mov r8,  [r12+8]    ; fd
+    mov r9,  [r12]      ; offset
+    add r12, 48         ; pop 6 args
+    mov rax, 9          ; syscall: mmap
+    syscall
+    mov rsp, rax        ; restore rsp
+    sub r12, 8
+    mov [r12], rax      ; push result
+    ret
     ret
 word_munmap:
 	mov rsi, [r12]
diff --git a/build/hello.o b/build/hello.o
index 5ba708d..e3e581a 100644
Binary files a/build/hello.o and b/build/hello.o differ
diff --git a/build/main.asm b/build/main.asm
index e644000..0305d97 100644
--- a/build/main.asm
+++ b/build/main.asm
@@ -18,6 +18,32 @@ _start:
     mov rdi, rax
     mov rax, 60
     syscall
+word_c_40:
+	mov rax, [r12]
+	movzx rax, byte [rax]
+	mov [r12], rax
+	ret
+    ret
+word_c_21:
+	mov rax, [r12]
+	add r12, 8
+	mov rbx, [r12]
+	mov [rbx], al
+	ret
+    ret
+word_r_40:
+	mov rax, [r13]
+	sub r12, 8
+	mov [r12], rax
+	ret
+    ret
+word_strlen:
+	mov rax, [r12]      ; addr
+	mov rcx, [r12 + 8]  ; len
+	add r12, 16         ; pop len and addr
+	mov [r12], rcx      ; push len
+	ret
+    ret
 word_puts:
 	; detects string if top is len>=0 and next is a pointer in [data_start, data_end]
 	mov rax, [r12]      ; len or int value
@@ -100,6 +126,15 @@ word_dup:
 	sub r12, 8
 	mov [r12], rax
     ret
+word_write_buf:
+	mov rdx, [r12]        ; len
+	mov rsi, [r12 + 8]    ; addr
+	add r12, 16           ; pop len + addr
+	mov rax, 1            ; syscall: write
+	mov rdi, 1            ; fd = stdout
+	syscall
+	ret
+    ret
 word_drop:
 	add r12, 8
     ret
@@ -271,20 +306,22 @@ word__21:
 	add r12, 8
     ret
 word_mmap:
-	mov r9, [r12]
-	add r12, 8
-	mov r8, [r12]
-	add r12, 8
-	mov r10, [r12]
-	add r12, 8
-	mov rdx, [r12]
-	add r12, 8
-	mov rsi, [r12]
-	add r12, 8
-	mov rdi, [r12]
-	mov rax, 9
-	syscall
-	mov [r12], rax
+    ; Save rsp and align to 16 bytes for syscall ABI
+    mov rax, rsp
+    and rsp, -16
+    mov rdi, [r12+40]   ; addr
+    mov rsi, [r12+32]   ; length
+    mov rdx, [r12+24]   ; prot
+    mov r10, [r12+16]   ; flags
+    mov r8,  [r12+8]    ; fd
+    mov r9,  [r12]      ; offset
+    add r12, 48         ; pop 6 args
+    mov rax, 9          ; syscall: mmap
+    syscall
+    mov rsp, rax        ; restore rsp
+    sub r12, 8
+    mov [r12], rax      ; push result
+    ret
     ret
 word_munmap:
 	mov rsi, [r12]
diff --git a/build/main.o b/build/main.o
index 5f778c4..e49159b 100644
Binary files a/build/main.o and b/build/main.o differ
diff --git a/build/str.asm b/build/str.asm
index aba2433..096670b 100644
--- a/build/str.asm
+++ b/build/str.asm
@@ -472,7 +472,57 @@ word_read_file:
 	mov qword [r12 + 8], -3
 	ret
     ret
-word_main:
+word_write_file:
+	; stack: path_ptr (top), path_len, buf_ptr, buf_len
+	mov rsi, [r12]        ; path_ptr
+	mov rdx, [r12 + 8]    ; path_len
+	mov r15, [r12 + 16]   ; buf_ptr (save in callee-saved r15)
+	mov r13, [r12 + 24]   ; buf_len (save in callee-saved r13)
+	add r12, 32           ; pop 4 args (we saved buf info)
+
+	; open(path_ptr, O_WRONLY|O_CREAT|O_TRUNC, 0666)
+	mov rdi, rsi          ; filename
+	mov rsi, 577          ; flags = O_WRONLY|O_CREAT|O_TRUNC
+	mov rdx, 438          ; mode = 0o666
+	mov rax, 2            ; syscall: open
+	syscall
+	cmp rax, 0
+	jl .fail_open
+	mov r9, rax           ; save fd
+
+	; write(fd, buf_ptr, buf_len) -- use preserved r15/r13 which survive syscalls
+	mov rax, 1            ; syscall: write
+	mov rdi, r9           ; fd
+	mov rsi, r15          ; buf_ptr
+	mov rdx, r13          ; buf_len
+	syscall
+	mov r10, rax          ; save write result
+	cmp r10, 0
+	jl .fail_write
+
+	; close(fd)
+	mov rax, 3            ; syscall: close
+	mov rdi, r9
+	syscall
+
+	sub r12, 8
+	mov [r12], r10
+	ret
+
+.fail_write:
+	mov rax, 3
+	mov rdi, r9
+	syscall
+	sub r12, 8
+	mov [r12], r10
+	ret
+
+.fail_open:
+	sub r12, 8
+	mov [r12], rax
+	ret
+    ret
+word_main2:
     ; push str_0
     sub r12, 8
     mov qword [r12], str_0
@@ -486,10 +536,22 @@ word_main:
     sub r12, 8
     mov qword [r12], 0
     ret
+word_main:
+    ; push str_1
+    sub r12, 8
+    mov qword [r12], str_1
+    ; push 7
+    sub r12, 8
+    mov qword [r12], 7
+    call word_dup
+    call word_puts
+    ret
 section .data
 data_start:
 str_0: db 115, 116, 114, 46, 115, 108, 0
 str_0_len equ 6
+str_1: db 118, 99, 120, 122, 118, 103, 103, 0
+str_1_len equ 7
 data_end:
 section .bss
 align 16
diff --git a/build/str.o b/build/str.o
index 5c58829..64e6be2 100644
Binary files a/build/str.o and b/build/str.o differ
diff --git a/build/test_write_file.asm b/build/test_write_file.asm
index 841bc91..a4fb632 100644
--- a/build/test_write_file.asm
+++ b/build/test_write_file.asm
@@ -526,18 +526,17 @@ word_main:
     ; push str_0
     sub r12, 8
     mov qword [r12], str_0
-    ; push 22
-    sub r12, 8
-    mov qword [r12], 22
-    call word_swap
-    ; push str_1
-    sub r12, 8
-    mov qword [r12], str_1
     ; push 27
     sub r12, 8
     mov qword [r12], 27
     call word_swap
-    call word_2swap
+    ; push str_1
+    sub r12, 8
+    mov qword [r12], str_1
+    ; push 22
+    sub r12, 8
+    mov qword [r12], 22
+    call word_swap
     call word_write_file
     call word_dup
     ; push 0
@@ -573,10 +572,10 @@ L_if_false_0:
     ret
 section .data
 data_start:
-str_0: db 47, 116, 109, 112, 47, 108, 50, 95, 116, 101, 115, 116, 95, 119, 114, 105, 116, 101, 46, 116, 120, 116, 0
-str_0_len equ 22
-str_1: db 104, 101, 108, 108, 111, 32, 102, 114, 111, 109, 32, 119, 114, 105, 116, 101, 95, 102, 105, 108, 101, 32, 116, 101, 115, 116, 10, 0
-str_1_len equ 27
+str_0: db 104, 101, 108, 108, 111, 32, 102, 114, 111, 109, 32, 119, 114, 105, 116, 101, 95, 102, 105, 108, 101, 32, 116, 101, 115, 116, 10, 0
+str_0_len equ 27
+str_1: db 47, 116, 109, 112, 47, 108, 50, 95, 116, 101, 115, 116, 95, 119, 114, 105, 116, 101, 46, 116, 120, 116, 0
+str_1_len equ 22
 str_2: db 119, 114, 111, 116, 101, 32, 98, 121, 116, 101, 115, 58, 32, 0
 str_2_len equ 13
 str_3: db 119, 114, 105, 116, 101, 32, 102, 97, 105, 108, 101, 100, 32, 101, 114, 114, 110, 111, 61, 0
diff --git a/build/test_write_file.o b/build/test_write_file.o
index b6082c5..1851e6d 100644
Binary files a/build/test_write_file.o and b/build/test_write_file.o differ
diff --git a/stdlib/stdlib.sl b/stdlib/stdlib.sl
index 37d4e7f..e39df4c 100644
--- a/stdlib/stdlib.sl
+++ b/stdlib/stdlib.sl
@@ -26,16 +26,6 @@
 }
 ;
 
-# : strlen ( addr len -- len )
-:asm strlen {
-	mov rax, [r12]      ; addr
-	mov rcx, [r12 + 8]  ; len
-	add r12, 16         ; pop len and addr
-	mov [r12], rcx      ; push len
-	ret
-}
-;
-
 :asm puts {
 	; detects string if top is len>=0 and next is a pointer in [data_start, data_end]
 	mov rax, [r12]      ; len or int value
diff --git a/test_write_file.sl b/test_write_file.sl
index 002f8f4..7e1c29a 100644
--- a/test_write_file.sl
+++ b/test_write_file.sl
@@ -2,11 +2,10 @@ import stdlib/stdlib.sl
 import stdlib/io.sl
 
 : main
-    "/tmp/l2_test_write.txt" # push path (len addr)
-    swap                      # -> (addr len) = path_ptr path_len
     "hello from write_file test\n" # push buf (len addr)
     swap                      # -> (addr len) = buf_ptr buf_len
-    2swap                     # reorder pairs -> path_ptr path_len buf_ptr buf_len
+    "/tmp/l2_test_write.txt" # push path (len addr)
+    swap                      # -> (addr len) = path_ptr path_len
     write_file
     dup 0 > if
         "wrote bytes: " puts
diff --git a/write b/write
index 7daac9d..ec3c604 100755
Binary files a/write and b/write differ